X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FVectorize%2FSLPVectorizer.cpp;h=fb1f64ba14ab39781afc30f260563989a8c96d56;hb=591e043a7af04d1d5e13beb22867fef32985c57b;hp=74a1ed69cd276b01acce3c791b1a66947755c1b0;hpb=e47fce60e4d34596c389b6ef64ca58b8749cb54e;p=oota-llvm.git diff --git a/lib/Transforms/Vectorize/SLPVectorizer.cpp b/lib/Transforms/Vectorize/SLPVectorizer.cpp index 74a1ed69cd2..fb1f64ba14a 100644 --- a/lib/Transforms/Vectorize/SLPVectorizer.cpp +++ b/lib/Transforms/Vectorize/SLPVectorizer.cpp @@ -19,7 +19,11 @@ #include "llvm/ADT/MapVector.h" #include "llvm/ADT/PostOrderIterator.h" #include "llvm/ADT/SetVector.h" +#include "llvm/ADT/Optional.h" +#include "llvm/ADT/Statistic.h" #include "llvm/Analysis/AliasAnalysis.h" +#include "llvm/Analysis/AssumptionCache.h" +#include "llvm/Analysis/CodeMetrics.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/ScalarEvolution.h" #include "llvm/Analysis/ScalarEvolutionExpressions.h" @@ -42,12 +46,15 @@ #include "llvm/Transforms/Utils/VectorUtils.h" #include #include +#include using namespace llvm; #define SV_NAME "slp-vectorizer" #define DEBUG_TYPE "SLP" +STATISTIC(NumVectorInstructions, "Number of vector instructions generated"); + static cl::opt SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden, cl::desc("Only vectorize if you gain more than this " @@ -68,52 +75,14 @@ static const unsigned MinVecRegSize = 128; static const unsigned RecursionMaxDepth = 12; -/// A helper class for numbering instructions in multiple blocks. -/// Numbers start at zero for each basic block. -struct BlockNumbering { - - BlockNumbering(BasicBlock *Bb) : BB(Bb), Valid(false) {} - - void numberInstructions() { - unsigned Loc = 0; - InstrIdx.clear(); - InstrVec.clear(); - // Number the instructions in the block. - for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { - InstrIdx[it] = Loc++; - InstrVec.push_back(it); - assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation"); - } - Valid = true; - } - - int getIndex(Instruction *I) { - assert(I->getParent() == BB && "Invalid instruction"); - if (!Valid) - numberInstructions(); - assert(InstrIdx.count(I) && "Unknown instruction"); - return InstrIdx[I]; - } - - Instruction *getInstruction(unsigned loc) { - if (!Valid) - numberInstructions(); - assert(InstrVec.size() > loc && "Invalid Index"); - return InstrVec[loc]; - } +// Limit the number of alias checks. The limit is chosen so that +// it has no negative effect on the llvm benchmarks. +static const unsigned AliasedCheckLimit = 10; - void forget() { Valid = false; } - -private: - /// The block we are numbering. - BasicBlock *BB; - /// Is the block numbered. - bool Valid; - /// Maps instructions to numbers and back. - SmallDenseMap InstrIdx; - /// Maps integers to Instructions. - SmallVector InstrVec; -}; +// Another limit for the alias checks: The maximum distance between load/store +// instructions where alias checks are done. +// This limit is useful for very large basic blocks. +static const unsigned MaxMemDepDistance = 160; /// \returns the parent basic block if all of the instructions in \p VL /// are in the same block or null otherwise. @@ -209,6 +178,23 @@ static unsigned getSameOpcode(ArrayRef VL) { return Opcode; } +/// Get the intersection (logical and) of all of the potential IR flags +/// of each scalar operation (VL) that will be converted into a vector (I). +/// Flag set: NSW, NUW, exact, and all of fast-math. +static void propagateIRFlags(Value *I, ArrayRef VL) { + if (auto *VecOp = dyn_cast(I)) { + if (auto *Intersection = dyn_cast(VL[0])) { + // Intersection is initialized to the 0th scalar, + // so start counting from index '1'. + for (int i = 1, e = VL.size(); i < e; ++i) { + if (auto *Scalar = dyn_cast(VL[i])) + Intersection->andIRFlags(Scalar); + } + VecOp->copyIRFlags(Intersection); + } + } +} + /// \returns \p I after propagating metadata from \p VL. static Instruction *propagateMetadata(Instruction *I, ArrayRef VL) { Instruction *I0 = cast(VL[0]); @@ -287,104 +273,42 @@ static bool CanReuseExtract(ArrayRef VL) { return true; } -static void reorderInputsAccordingToOpcode(ArrayRef VL, - SmallVectorImpl &Left, - SmallVectorImpl &Right) { - - SmallVector OrigLeft, OrigRight; +/// \returns True if in-tree use also needs extract. This refers to +/// possible scalar operand in vectorized instruction. +static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst, + TargetLibraryInfo *TLI) { - bool AllSameOpcodeLeft = true; - bool AllSameOpcodeRight = true; - for (unsigned i = 0, e = VL.size(); i != e; ++i) { - Instruction *I = cast(VL[i]); - Value *V0 = I->getOperand(0); - Value *V1 = I->getOperand(1); - - OrigLeft.push_back(V0); - OrigRight.push_back(V1); - - Instruction *I0 = dyn_cast(V0); - Instruction *I1 = dyn_cast(V1); - - // Check whether all operands on one side have the same opcode. In this case - // we want to preserve the original order and not make things worse by - // reordering. - AllSameOpcodeLeft = I0; - AllSameOpcodeRight = I1; - - if (i && AllSameOpcodeLeft) { - if(Instruction *P0 = dyn_cast(OrigLeft[i-1])) { - if(P0->getOpcode() != I0->getOpcode()) - AllSameOpcodeLeft = false; - } else - AllSameOpcodeLeft = false; - } - if (i && AllSameOpcodeRight) { - if(Instruction *P1 = dyn_cast(OrigRight[i-1])) { - if(P1->getOpcode() != I1->getOpcode()) - AllSameOpcodeRight = false; - } else - AllSameOpcodeRight = false; - } - - // Sort two opcodes. In the code below we try to preserve the ability to use - // broadcast of values instead of individual inserts. - // vl1 = load - // vl2 = phi - // vr1 = load - // vr2 = vr2 - // = vl1 x vr1 - // = vl2 x vr2 - // If we just sorted according to opcode we would leave the first line in - // tact but we would swap vl2 with vr2 because opcode(phi) > opcode(load). - // = vl1 x vr1 - // = vr2 x vl2 - // Because vr2 and vr1 are from the same load we loose the opportunity of a - // broadcast for the packed right side in the backend: we have [vr1, vl2] - // instead of [vr1, vr2=vr1]. - if (I0 && I1) { - if(!i && I0->getOpcode() > I1->getOpcode()) { - Left.push_back(I1); - Right.push_back(I0); - } else if (i && I0->getOpcode() > I1->getOpcode() && Right[i-1] != I1) { - // Try not to destroy a broad cast for no apparent benefit. - Left.push_back(I1); - Right.push_back(I0); - } else if (i && I0->getOpcode() == I1->getOpcode() && Right[i-1] == I0) { - // Try preserve broadcasts. - Left.push_back(I1); - Right.push_back(I0); - } else if (i && I0->getOpcode() == I1->getOpcode() && Left[i-1] == I1) { - // Try preserve broadcasts. - Left.push_back(I1); - Right.push_back(I0); - } else { - Left.push_back(I0); - Right.push_back(I1); - } - continue; - } - // One opcode, put the instruction on the right. - if (I0) { - Left.push_back(V1); - Right.push_back(I0); - continue; + unsigned Opcode = UserInst->getOpcode(); + switch (Opcode) { + case Instruction::Load: { + LoadInst *LI = cast(UserInst); + return (LI->getPointerOperand() == Scalar); + } + case Instruction::Store: { + StoreInst *SI = cast(UserInst); + return (SI->getPointerOperand() == Scalar); + } + case Instruction::Call: { + CallInst *CI = cast(UserInst); + Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); + if (hasVectorInstrinsicScalarOpd(ID, 1)) { + return (CI->getArgOperand(1) == Scalar); } - Left.push_back(V0); - Right.push_back(V1); } - - bool LeftBroadcast = isSplat(Left); - bool RightBroadcast = isSplat(Right); - - // Don't reorder if the operands where good to begin with. - if (!(LeftBroadcast || RightBroadcast) && - (AllSameOpcodeRight || AllSameOpcodeLeft)) { - Left = OrigLeft; - Right = OrigRight; + default: + return false; } } +/// \returns the AA location that is being access by the instruction. +static AliasAnalysis::Location getLocation(Instruction *I, AliasAnalysis *AA) { + if (StoreInst *SI = dyn_cast(I)) + return AA->getLocation(SI); + if (LoadInst *LI = dyn_cast(I)) + return AA->getLocation(LI); + return AliasAnalysis::Location(); +} + /// Bottom Up SLP Vectorizer. class BoUpSLP { public: @@ -395,15 +319,21 @@ public: BoUpSLP(Function *Func, ScalarEvolution *Se, const DataLayout *Dl, TargetTransformInfo *Tti, TargetLibraryInfo *TLi, AliasAnalysis *Aa, - LoopInfo *Li, DominatorTree *Dt) - : NumLoadsWantToKeepOrder(0), NumLoadsWantToChangeOrder(0), - F(Func), SE(Se), DL(Dl), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt), - Builder(Se->getContext()) {} + LoopInfo *Li, DominatorTree *Dt, AssumptionCache *AC) + : NumLoadsWantToKeepOrder(0), NumLoadsWantToChangeOrder(0), F(Func), + SE(Se), DL(Dl), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt), + Builder(Se->getContext()) { + CodeMetrics::collectEphemeralValues(F, AC, EphValues); + } /// \brief Vectorize the tree that starts with the elements in \p VL. /// Returns the vectorized root. Value *vectorizeTree(); + /// \returns the cost incurred by unwanted spills and fills, caused by + /// holding live values over call sites. + int getSpillCost(); + /// \returns the vectorization cost of the subtree that starts at \p VL. /// A negative number means that this is profitable. int getTreeCost(); @@ -419,9 +349,12 @@ public: ScalarToTreeEntry.clear(); MustGather.clear(); ExternalUses.clear(); - MemBarrierIgnoreList.clear(); NumLoadsWantToKeepOrder = 0; NumLoadsWantToChangeOrder = 0; + for (auto &Iter : BlocksSchedules) { + BlockScheduling *BS = Iter.second.get(); + BS->clear(); + } } /// \returns true if the memory operations A and B are consecutive. @@ -471,20 +404,6 @@ private: /// roots. This method calculates the cost of extracting the values. int getGatherCost(ArrayRef VL); - /// \returns the AA location that is being access by the instruction. - AliasAnalysis::Location getLocation(Instruction *I); - - /// \brief Checks if it is possible to sink an instruction from - /// \p Src to \p Dst. - /// \returns the pointer to the barrier instruction if we can't sink. - Value *getSinkBarrier(Instruction *Src, Instruction *Dst); - - /// \returns the index of the last instruction in the BB from \p VL. - int getLastIndex(ArrayRef VL); - - /// \returns the Instruction in the bundle \p VL. - Instruction *getLastInstruction(ArrayRef VL); - /// \brief Set the Builder insert point to one after the last instruction in /// the bundle void setInsertPointAfterBundle(ArrayRef VL); @@ -496,8 +415,18 @@ private: /// be beneficial even the tree height is tiny. bool isFullyVectorizableTinyTree(); + /// \reorder commutative operands in alt shuffle if they result in + /// vectorized code. + void reorderAltShuffleOperands(ArrayRef VL, + SmallVectorImpl &Left, + SmallVectorImpl &Right); + /// \reorder commutative operands to get better probability of + /// generating vectorized code. + void reorderInputsAccordingToOpcode(ArrayRef VL, + SmallVectorImpl &Left, + SmallVectorImpl &Right); struct TreeEntry { - TreeEntry() : Scalars(), VectorizedValue(nullptr), LastScalarIndex(0), + TreeEntry() : Scalars(), VectorizedValue(nullptr), NeedToGather(0) {} /// \returns true if the scalars in VL are equal to this entry. @@ -512,9 +441,6 @@ private: /// The Scalars are vectorized into this value. It is initialized to Null. Value *VectorizedValue; - /// The index in the basic block of the last scalar. - int LastScalarIndex; - /// Do we need to gather this sequence ? bool NeedToGather; }; @@ -527,18 +453,16 @@ private: Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end()); Last->NeedToGather = !Vectorized; if (Vectorized) { - Last->LastScalarIndex = getLastIndex(VL); for (int i = 0, e = VL.size(); i != e; ++i) { assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!"); ScalarToTreeEntry[VL[i]] = idx; } } else { - Last->LastScalarIndex = 0; MustGather.insert(VL.begin(), VL.end()); } return Last; } - + /// -- Vectorization State -- /// Holds all of the tree entries. std::vector VectorizableTree; @@ -562,28 +486,359 @@ private: }; typedef SmallVector UserList; + /// Checks if two instructions may access the same memory. + /// + /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it + /// is invariant in the calling loop. + bool isAliased(const AliasAnalysis::Location &Loc1, Instruction *Inst1, + Instruction *Inst2) { + + // First check if the result is already in the cache. + AliasCacheKey key = std::make_pair(Inst1, Inst2); + Optional &result = AliasCache[key]; + if (result.hasValue()) { + return result.getValue(); + } + AliasAnalysis::Location Loc2 = getLocation(Inst2, AA); + bool aliased = true; + if (Loc1.Ptr && Loc2.Ptr) { + // Do the alias check. + aliased = AA->alias(Loc1, Loc2); + } + // Store the result in the cache. + result = aliased; + return aliased; + } + + typedef std::pair AliasCacheKey; + + /// Cache for alias results. + /// TODO: consider moving this to the AliasAnalysis itself. + DenseMap> AliasCache; + + /// Removes an instruction from its block and eventually deletes it. + /// It's like Instruction::eraseFromParent() except that the actual deletion + /// is delayed until BoUpSLP is destructed. + /// This is required to ensure that there are no incorrect collisions in the + /// AliasCache, which can happen if a new instruction is allocated at the + /// same address as a previously deleted instruction. + void eraseInstruction(Instruction *I) { + I->removeFromParent(); + I->dropAllReferences(); + DeletedInstructions.push_back(std::unique_ptr(I)); + } + + /// Temporary store for deleted instructions. Instructions will be deleted + /// eventually when the BoUpSLP is destructed. + SmallVector, 8> DeletedInstructions; + /// A list of values that need to extracted out of the tree. /// This list holds pairs of (Internal Scalar : External User). UserList ExternalUses; - /// A list of instructions to ignore while sinking - /// memory instructions. This map must be reset between runs of getCost. - ValueSet MemBarrierIgnoreList; + /// Values used only by @llvm.assume calls. + SmallPtrSet EphValues; /// Holds all of the instructions that we gathered. SetVector GatherSeq; /// A list of blocks that we are going to CSE. SetVector CSEBlocks; - /// Numbers instructions in different blocks. - DenseMap BlocksNumbers; + /// Contains all scheduling relevant data for an instruction. + /// A ScheduleData either represents a single instruction or a member of an + /// instruction bundle (= a group of instructions which is combined into a + /// vector instruction). + struct ScheduleData { + + // The initial value for the dependency counters. It means that the + // dependencies are not calculated yet. + enum { InvalidDeps = -1 }; + + ScheduleData() + : Inst(nullptr), FirstInBundle(nullptr), NextInBundle(nullptr), + NextLoadStore(nullptr), SchedulingRegionID(0), SchedulingPriority(0), + Dependencies(InvalidDeps), UnscheduledDeps(InvalidDeps), + UnscheduledDepsInBundle(InvalidDeps), IsScheduled(false) {} + + void init(int BlockSchedulingRegionID) { + FirstInBundle = this; + NextInBundle = nullptr; + NextLoadStore = nullptr; + IsScheduled = false; + SchedulingRegionID = BlockSchedulingRegionID; + UnscheduledDepsInBundle = UnscheduledDeps; + clearDependencies(); + } + + /// Returns true if the dependency information has been calculated. + bool hasValidDependencies() const { return Dependencies != InvalidDeps; } + + /// Returns true for single instructions and for bundle representatives + /// (= the head of a bundle). + bool isSchedulingEntity() const { return FirstInBundle == this; } + + /// Returns true if it represents an instruction bundle and not only a + /// single instruction. + bool isPartOfBundle() const { + return NextInBundle != nullptr || FirstInBundle != this; + } + + /// Returns true if it is ready for scheduling, i.e. it has no more + /// unscheduled depending instructions/bundles. + bool isReady() const { + assert(isSchedulingEntity() && + "can't consider non-scheduling entity for ready list"); + return UnscheduledDepsInBundle == 0 && !IsScheduled; + } + + /// Modifies the number of unscheduled dependencies, also updating it for + /// the whole bundle. + int incrementUnscheduledDeps(int Incr) { + UnscheduledDeps += Incr; + return FirstInBundle->UnscheduledDepsInBundle += Incr; + } + + /// Sets the number of unscheduled dependencies to the number of + /// dependencies. + void resetUnscheduledDeps() { + incrementUnscheduledDeps(Dependencies - UnscheduledDeps); + } + + /// Clears all dependency information. + void clearDependencies() { + Dependencies = InvalidDeps; + resetUnscheduledDeps(); + MemoryDependencies.clear(); + } + + void dump(raw_ostream &os) const { + if (!isSchedulingEntity()) { + os << "/ " << *Inst; + } else if (NextInBundle) { + os << '[' << *Inst; + ScheduleData *SD = NextInBundle; + while (SD) { + os << ';' << *SD->Inst; + SD = SD->NextInBundle; + } + os << ']'; + } else { + os << *Inst; + } + } + + Instruction *Inst; - /// \brief Get the corresponding instruction numbering list for a given - /// BasicBlock. The list is allocated lazily. - BlockNumbering &getBlockNumbering(BasicBlock *BB) { - auto I = BlocksNumbers.insert(std::make_pair(BB, BlockNumbering(BB))); - return I.first->second; - } + /// Points to the head in an instruction bundle (and always to this for + /// single instructions). + ScheduleData *FirstInBundle; + + /// Single linked list of all instructions in a bundle. Null if it is a + /// single instruction. + ScheduleData *NextInBundle; + + /// Single linked list of all memory instructions (e.g. load, store, call) + /// in the block - until the end of the scheduling region. + ScheduleData *NextLoadStore; + + /// The dependent memory instructions. + /// This list is derived on demand in calculateDependencies(). + SmallVector MemoryDependencies; + + /// This ScheduleData is in the current scheduling region if this matches + /// the current SchedulingRegionID of BlockScheduling. + int SchedulingRegionID; + + /// Used for getting a "good" final ordering of instructions. + int SchedulingPriority; + + /// The number of dependencies. Constitutes of the number of users of the + /// instruction plus the number of dependent memory instructions (if any). + /// This value is calculated on demand. + /// If InvalidDeps, the number of dependencies is not calculated yet. + /// + int Dependencies; + + /// The number of dependencies minus the number of dependencies of scheduled + /// instructions. As soon as this is zero, the instruction/bundle gets ready + /// for scheduling. + /// Note that this is negative as long as Dependencies is not calculated. + int UnscheduledDeps; + + /// The sum of UnscheduledDeps in a bundle. Equals to UnscheduledDeps for + /// single instructions. + int UnscheduledDepsInBundle; + + /// True if this instruction is scheduled (or considered as scheduled in the + /// dry-run). + bool IsScheduled; + }; + +#ifndef NDEBUG + friend raw_ostream &operator<<(raw_ostream &os, + const BoUpSLP::ScheduleData &SD); +#endif + + /// Contains all scheduling data for a basic block. + /// + struct BlockScheduling { + + BlockScheduling(BasicBlock *BB) + : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize), + ScheduleStart(nullptr), ScheduleEnd(nullptr), + FirstLoadStoreInRegion(nullptr), LastLoadStoreInRegion(nullptr), + // Make sure that the initial SchedulingRegionID is greater than the + // initial SchedulingRegionID in ScheduleData (which is 0). + SchedulingRegionID(1) {} + + void clear() { + ReadyInsts.clear(); + ScheduleStart = nullptr; + ScheduleEnd = nullptr; + FirstLoadStoreInRegion = nullptr; + LastLoadStoreInRegion = nullptr; + + // Make a new scheduling region, i.e. all existing ScheduleData is not + // in the new region yet. + ++SchedulingRegionID; + } + + ScheduleData *getScheduleData(Value *V) { + ScheduleData *SD = ScheduleDataMap[V]; + if (SD && SD->SchedulingRegionID == SchedulingRegionID) + return SD; + return nullptr; + } + + bool isInSchedulingRegion(ScheduleData *SD) { + return SD->SchedulingRegionID == SchedulingRegionID; + } + + /// Marks an instruction as scheduled and puts all dependent ready + /// instructions into the ready-list. + template + void schedule(ScheduleData *SD, ReadyListType &ReadyList) { + SD->IsScheduled = true; + DEBUG(dbgs() << "SLP: schedule " << *SD << "\n"); + + ScheduleData *BundleMember = SD; + while (BundleMember) { + // Handle the def-use chain dependencies. + for (Use &U : BundleMember->Inst->operands()) { + ScheduleData *OpDef = getScheduleData(U.get()); + if (OpDef && OpDef->hasValidDependencies() && + OpDef->incrementUnscheduledDeps(-1) == 0) { + // There are no more unscheduled dependencies after decrementing, + // so we can put the dependent instruction into the ready list. + ScheduleData *DepBundle = OpDef->FirstInBundle; + assert(!DepBundle->IsScheduled && + "already scheduled bundle gets ready"); + ReadyList.insert(DepBundle); + DEBUG(dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"); + } + } + // Handle the memory dependencies. + for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) { + if (MemoryDepSD->incrementUnscheduledDeps(-1) == 0) { + // There are no more unscheduled dependencies after decrementing, + // so we can put the dependent instruction into the ready list. + ScheduleData *DepBundle = MemoryDepSD->FirstInBundle; + assert(!DepBundle->IsScheduled && + "already scheduled bundle gets ready"); + ReadyList.insert(DepBundle); + DEBUG(dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"); + } + } + BundleMember = BundleMember->NextInBundle; + } + } + + /// Put all instructions into the ReadyList which are ready for scheduling. + template + void initialFillReadyList(ReadyListType &ReadyList) { + for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { + ScheduleData *SD = getScheduleData(I); + if (SD->isSchedulingEntity() && SD->isReady()) { + ReadyList.insert(SD); + DEBUG(dbgs() << "SLP: initially in ready list: " << *I << "\n"); + } + } + } + + /// Checks if a bundle of instructions can be scheduled, i.e. has no + /// cyclic dependencies. This is only a dry-run, no instructions are + /// actually moved at this stage. + bool tryScheduleBundle(ArrayRef VL, BoUpSLP *SLP); + + /// Un-bundles a group of instructions. + void cancelScheduling(ArrayRef VL); + + /// Extends the scheduling region so that V is inside the region. + void extendSchedulingRegion(Value *V); + + /// Initialize the ScheduleData structures for new instructions in the + /// scheduling region. + void initScheduleData(Instruction *FromI, Instruction *ToI, + ScheduleData *PrevLoadStore, + ScheduleData *NextLoadStore); + + /// Updates the dependency information of a bundle and of all instructions/ + /// bundles which depend on the original bundle. + void calculateDependencies(ScheduleData *SD, bool InsertInReadyList, + BoUpSLP *SLP); + + /// Sets all instruction in the scheduling region to un-scheduled. + void resetSchedule(); + + BasicBlock *BB; + + /// Simple memory allocation for ScheduleData. + std::vector> ScheduleDataChunks; + + /// The size of a ScheduleData array in ScheduleDataChunks. + int ChunkSize; + + /// The allocator position in the current chunk, which is the last entry + /// of ScheduleDataChunks. + int ChunkPos; + + /// Attaches ScheduleData to Instruction. + /// Note that the mapping survives during all vectorization iterations, i.e. + /// ScheduleData structures are recycled. + DenseMap ScheduleDataMap; + + struct ReadyList : SmallVector { + void insert(ScheduleData *SD) { push_back(SD); } + }; + + /// The ready-list for scheduling (only used for the dry-run). + ReadyList ReadyInsts; + + /// The first instruction of the scheduling region. + Instruction *ScheduleStart; + + /// The first instruction _after_ the scheduling region. + Instruction *ScheduleEnd; + + /// The first memory accessing instruction in the scheduling region + /// (can be null). + ScheduleData *FirstLoadStoreInRegion; + + /// The last memory accessing instruction in the scheduling region + /// (can be null). + ScheduleData *LastLoadStoreInRegion; + + /// The ID of the scheduling region. For a new vectorization iteration this + /// is incremented which "removes" all ScheduleData from the region. + int SchedulingRegionID; + }; + + /// Attaches the BlockScheduling structures to basic blocks. + MapVector> BlocksSchedules; + + /// Performs the "real" scheduling. Done before vectorization is actually + /// performed in a basic block. + void scheduleBlock(BlockScheduling *BS); /// List of users to ignore during scheduling and that don't need extracting. ArrayRef UserIgnoreList; @@ -607,6 +862,13 @@ private: IRBuilder<> Builder; }; +#ifndef NDEBUG +raw_ostream &operator<<(raw_ostream &os, const BoUpSLP::ScheduleData &SD) { + SD.dump(os); + return os; +} +#endif + void BoUpSLP::buildTree(ArrayRef Roots, ArrayRef UserIgnoreLst) { deleteTree(); @@ -630,18 +892,27 @@ void BoUpSLP::buildTree(ArrayRef Roots, for (User *U : Scalar->users()) { DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n"); - // Skip in-tree scalars that become vectors. - if (ScalarToTreeEntry.count(U)) { - DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << - *U << ".\n"); - int Idx = ScalarToTreeEntry[U]; (void) Idx; - assert(!VectorizableTree[Idx].NeedToGather && "Bad state"); - continue; - } Instruction *UserInst = dyn_cast(U); if (!UserInst) continue; + // Skip in-tree scalars that become vectors + if (ScalarToTreeEntry.count(U)) { + int Idx = ScalarToTreeEntry[U]; + TreeEntry *UseEntry = &VectorizableTree[Idx]; + Value *UseScalar = UseEntry->Scalars[0]; + // Some in-tree scalars will remain as scalar in vectorized + // instructions. If that is the case, the one in Lane 0 will + // be used. + if (UseScalar != U || + !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) { + DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << *U + << ".\n"); + assert(!VectorizableTree[Idx].NeedToGather && "Bad state"); + continue; + } + } + // Ignore users in the user ignore list. if (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), UserInst) != UserIgnoreList.end()) @@ -701,6 +972,16 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { // We now know that this is a vector of instructions of the same type from // the same block. + // Don't vectorize ephemeral values. + for (unsigned i = 0, e = VL.size(); i != e; ++i) { + if (EphValues.count(VL[i])) { + DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] << + ") is ephemeral.\n"); + newTreeEntry(VL, false); + return; + } + } + // Check if this is a duplicate of another entry. if (ScalarToTreeEntry.count(VL[0])) { int Idx = ScalarToTreeEntry[VL[0]]; @@ -727,11 +1008,11 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { } } - // If any of the scalars appears in the table OR it is marked as a value that - // needs to stat scalar then we need to gather the scalars. + // If any of the scalars is marked as a value that needs to stay scalar then + // we need to gather the scalars. for (unsigned i = 0, e = VL.size(); i != e; ++i) { - if (ScalarToTreeEntry.count(VL[i]) || MustGather.count(VL[i])) { - DEBUG(dbgs() << "SLP: Gathering due to gathered scalar. \n"); + if (MustGather.count(VL[i])) { + DEBUG(dbgs() << "SLP: Gathering due to gathered scalar.\n"); newTreeEntry(VL, false); return; } @@ -740,69 +1021,16 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { // Check that all of the users of the scalars that we want to vectorize are // schedulable. Instruction *VL0 = cast(VL[0]); - int MyLastIndex = getLastIndex(VL); BasicBlock *BB = cast(VL0)->getParent(); - for (unsigned i = 0, e = VL.size(); i != e; ++i) { - Instruction *Scalar = cast(VL[i]); - DEBUG(dbgs() << "SLP: Checking users of " << *Scalar << ". \n"); - for (User *U : Scalar->users()) { - DEBUG(dbgs() << "SLP: \tUser " << *U << ". \n"); - Instruction *UI = dyn_cast(U); - if (!UI) { - DEBUG(dbgs() << "SLP: Gathering due unknown user. \n"); - newTreeEntry(VL, false); - return; - } - - // We don't care if the user is in a different basic block. - BasicBlock *UserBlock = UI->getParent(); - if (UserBlock != BB) { - DEBUG(dbgs() << "SLP: User from a different basic block " - << *UI << ". \n"); - continue; - } - - // If this is a PHINode within this basic block then we can place the - // extract wherever we want. - if (isa(*UI)) { - DEBUG(dbgs() << "SLP: \tWe can schedule PHIs:" << *UI << ". \n"); - continue; - } - - // Check if this is a safe in-tree user. - if (ScalarToTreeEntry.count(UI)) { - int Idx = ScalarToTreeEntry[UI]; - int VecLocation = VectorizableTree[Idx].LastScalarIndex; - if (VecLocation <= MyLastIndex) { - DEBUG(dbgs() << "SLP: Gathering due to unschedulable vector. \n"); - newTreeEntry(VL, false); - return; - } - DEBUG(dbgs() << "SLP: In-tree user (" << *UI << ") at #" << - VecLocation << " vector value (" << *Scalar << ") at #" - << MyLastIndex << ".\n"); - continue; - } - - // Ignore users in the user ignore list. - if (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), UI) != - UserIgnoreList.end()) - continue; - - // Make sure that we can schedule this unknown user. - BlockNumbering &BN = getBlockNumbering(BB); - int UserIndex = BN.getIndex(UI); - if (UserIndex < MyLastIndex) { - - DEBUG(dbgs() << "SLP: Can't schedule extractelement for " - << *UI << ". \n"); - newTreeEntry(VL, false); - return; - } - } + if (!DT->isReachableFromEntry(BB)) { + // Don't go into unreachable blocks. They may contain instructions with + // dependency cycles which confuse the final scheduling. + DEBUG(dbgs() << "SLP: bundle in unreachable block.\n"); + newTreeEntry(VL, false); + return; } - + // Check that every instructions appears once in this bundle. for (unsigned i = 0, e = VL.size(); i < e; ++i) for (unsigned j = i+1; j < e; ++j) @@ -812,38 +1040,19 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { return; } - // Check that instructions in this bundle don't reference other instructions. - // The runtime of this check is O(N * N-1 * uses(N)) and a typical N is 4. - for (unsigned i = 0, e = VL.size(); i < e; ++i) { - for (User *U : VL[i]->users()) { - for (unsigned j = 0; j < e; ++j) { - if (i != j && U == VL[j]) { - DEBUG(dbgs() << "SLP: Intra-bundle dependencies!" << *U << ". \n"); - newTreeEntry(VL, false); - return; - } - } - } + auto &BSRef = BlocksSchedules[BB]; + if (!BSRef) { + BSRef = llvm::make_unique(BB); } + BlockScheduling &BS = *BSRef.get(); - DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n"); - - // Check if it is safe to sink the loads or the stores. - if (Opcode == Instruction::Load || Opcode == Instruction::Store) { - Instruction *Last = getLastInstruction(VL); - - for (unsigned i = 0, e = VL.size(); i < e; ++i) { - if (VL[i] == Last) - continue; - Value *Barrier = getSinkBarrier(cast(VL[i]), Last); - if (Barrier) { - DEBUG(dbgs() << "SLP: Can't sink " << *VL[i] << "\n down to " << *Last - << "\n because of " << *Barrier << ". Gathering.\n"); - newTreeEntry(VL, false); - return; - } - } + if (!BS.tryScheduleBundle(VL, this)) { + DEBUG(dbgs() << "SLP: We are not able to schedule this bundle!\n"); + BS.cancelScheduling(VL); + newTreeEntry(VL, false); + return; } + DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n"); switch (Opcode) { case Instruction::PHI: { @@ -856,6 +1065,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { cast(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i))); if (Term) { DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n"); + BS.cancelScheduling(VL); newTreeEntry(VL, false); return; } @@ -879,6 +1089,8 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { bool Reuse = CanReuseExtract(VL); if (Reuse) { DEBUG(dbgs() << "SLP: Reusing extract sequence.\n"); + } else { + BS.cancelScheduling(VL); } newTreeEntry(VL, Reuse); return; @@ -888,6 +1100,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) { LoadInst *L = cast(VL[i]); if (!L->isSimple()) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n"); return; @@ -896,6 +1109,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { if (VL.size() == 2 && isConsecutiveAccess(VL[1], VL[0])) { ++NumLoadsWantToChangeOrder; } + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n"); return; @@ -922,6 +1136,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { for (unsigned i = 0; i < VL.size(); ++i) { Type *Ty = cast(VL[i])->getOperand(0)->getType(); if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n"); return; @@ -949,6 +1164,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { CmpInst *Cmp = cast(VL[i]); if (Cmp->getPredicate() != P0 || Cmp->getOperand(0)->getType() != ComparedTy) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n"); return; @@ -995,20 +1211,8 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { if (isa(VL0) && VL0->isCommutative()) { ValueList Left, Right; reorderInputsAccordingToOpcode(VL, Left, Right); - BasicBlock *LeftBB = getSameBlock(Left); - BasicBlock *RightBB = getSameBlock(Right); - // If we have common uses on separate paths in the tree make sure we - // process the one with greater common depth first. - // We can use block numbering to determine the subtree traversal as - // earler user has to come in between the common use and the later user. - if (LeftBB && RightBB && LeftBB == RightBB && - getLastIndex(Right) > getLastIndex(Left)) { - buildTree_rec(Right, Depth + 1); - buildTree_rec(Left, Depth + 1); - } else { - buildTree_rec(Left, Depth + 1); - buildTree_rec(Right, Depth + 1); - } + buildTree_rec(Left, Depth + 1); + buildTree_rec(Right, Depth + 1); return; } @@ -1027,6 +1231,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { for (unsigned j = 0; j < VL.size(); ++j) { if (cast(VL[j])->getNumOperands() != 2) { DEBUG(dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n"); + BS.cancelScheduling(VL); newTreeEntry(VL, false); return; } @@ -1039,6 +1244,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { Type *CurTy = cast(VL[j])->getOperand(0)->getType(); if (Ty0 != CurTy) { DEBUG(dbgs() << "SLP: not-vectorizable GEP (different types).\n"); + BS.cancelScheduling(VL); newTreeEntry(VL, false); return; } @@ -1050,6 +1256,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { if (!isa(Op)) { DEBUG( dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n"); + BS.cancelScheduling(VL); newTreeEntry(VL, false); return; } @@ -1071,6 +1278,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { // Check if the stores are consecutive or of we need to swizzle them. for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) if (!isConsecutiveAccess(VL[i], VL[i + 1])) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Non-consecutive store.\n"); return; @@ -1083,8 +1291,6 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { for (unsigned j = 0; j < VL.size(); ++j) Operands.push_back(cast(VL[j])->getOperand(0)); - // We can ignore these values because we are sinking them down. - MemBarrierIgnoreList.insert(VL.begin(), VL.end()); buildTree_rec(Operands, Depth + 1); return; } @@ -1095,6 +1301,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { // represented by an intrinsic call Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI); if (!isTriviallyVectorizable(ID)) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Non-vectorizable call.\n"); return; @@ -1107,6 +1314,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { CallInst *CI2 = dyn_cast(VL[i]); if (!CI2 || CI2->getCalledFunction() != Int || getIntrinsicIDForCall(CI2, TLI) != ID) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *VL[i] << "\n"); @@ -1117,6 +1325,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { if (hasVectorInstrinsicScalarOpd(ID, 1)) { Value *A1J = CI2->getArgOperand(1); if (A1I != A1J) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument "<< A1I<<"!=" << A1J @@ -1142,12 +1351,23 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { // If this is not an alternate sequence of opcode like add-sub // then do not vectorize this instruction. if (!isAltShuffle) { + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n"); return; } newTreeEntry(VL, true); DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n"); + + // Reorder operands if reordering would enable vectorization. + if (isa(VL0)) { + ValueList Left, Right; + reorderAltShuffleOperands(VL, Left, Right); + buildTree_rec(Left, Depth + 1); + buildTree_rec(Right, Depth + 1); + return; + } + for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { ValueList Operands; // Prepare the operand vector. @@ -1159,6 +1379,7 @@ void BoUpSLP::buildTree_rec(ArrayRef VL, unsigned Depth) { return; } default: + BS.cancelScheduling(VL); newTreeEntry(VL, false); DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n"); return; @@ -1261,6 +1482,10 @@ int BoUpSLP::getEntryCost(TreeEntry *E) { TargetTransformInfo::OK_AnyValue; TargetTransformInfo::OperandValueKind Op2VK = TargetTransformInfo::OK_UniformConstantValue; + TargetTransformInfo::OperandValueProperties Op1VP = + TargetTransformInfo::OP_None; + TargetTransformInfo::OperandValueProperties Op2VP = + TargetTransformInfo::OP_None; // If all operands are exactly the same ConstantInt then set the // operand kind to OK_UniformConstantValue. @@ -1282,11 +1507,17 @@ int BoUpSLP::getEntryCost(TreeEntry *E) { CInt != cast(I->getOperand(1))) Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; } + // FIXME: Currently cost of model modification for division by + // power of 2 is handled only for X86. Add support for other targets. + if (Op2VK == TargetTransformInfo::OK_UniformConstantValue && CInt && + CInt->getValue().isPowerOf2()) + Op2VP = TargetTransformInfo::OP_PowerOf2; - ScalarCost = - VecTy->getNumElements() * - TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK); - VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK); + ScalarCost = VecTy->getNumElements() * + TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK, + Op1VP, Op2VP); + VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK, + Op1VP, Op2VP); } return VecCost - ScalarCost; } @@ -1391,6 +1622,68 @@ bool BoUpSLP::isFullyVectorizableTinyTree() { return true; } +int BoUpSLP::getSpillCost() { + // Walk from the bottom of the tree to the top, tracking which values are + // live. When we see a call instruction that is not part of our tree, + // query TTI to see if there is a cost to keeping values live over it + // (for example, if spills and fills are required). + unsigned BundleWidth = VectorizableTree.front().Scalars.size(); + int Cost = 0; + + SmallPtrSet LiveValues; + Instruction *PrevInst = nullptr; + + for (unsigned N = 0; N < VectorizableTree.size(); ++N) { + Instruction *Inst = dyn_cast(VectorizableTree[N].Scalars[0]); + if (!Inst) + continue; + + if (!PrevInst) { + PrevInst = Inst; + continue; + } + + DEBUG( + dbgs() << "SLP: #LV: " << LiveValues.size(); + for (auto *X : LiveValues) + dbgs() << " " << X->getName(); + dbgs() << ", Looking at "; + Inst->dump(); + ); + + // Update LiveValues. + LiveValues.erase(PrevInst); + for (auto &J : PrevInst->operands()) { + if (isa(&*J) && ScalarToTreeEntry.count(&*J)) + LiveValues.insert(cast(&*J)); + } + + // Now find the sequence of instructions between PrevInst and Inst. + BasicBlock::reverse_iterator InstIt(Inst), PrevInstIt(PrevInst); + --PrevInstIt; + while (InstIt != PrevInstIt) { + if (PrevInstIt == PrevInst->getParent()->rend()) { + PrevInstIt = Inst->getParent()->rbegin(); + continue; + } + + if (isa(&*PrevInstIt) && &*PrevInstIt != PrevInst) { + SmallVector V; + for (auto *II : LiveValues) + V.push_back(VectorType::get(II->getType(), BundleWidth)); + Cost += TTI->getCostOfKeepingLiveOverCall(V); + } + + ++PrevInstIt; + } + + PrevInst = Inst; + } + + DEBUG(dbgs() << "SLP: SpillCost=" << Cost << "\n"); + return Cost; +} + int BoUpSLP::getTreeCost() { int Cost = 0; DEBUG(dbgs() << "SLP: Calculating cost for tree of size " << @@ -1398,7 +1691,7 @@ int BoUpSLP::getTreeCost() { // We only vectorize tiny trees if it is fully vectorizable. if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) { - if (!VectorizableTree.size()) { + if (VectorizableTree.empty()) { assert(!ExternalUses.size() && "We should not have any external users"); } return INT_MAX; @@ -1418,7 +1711,13 @@ int BoUpSLP::getTreeCost() { for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end(); I != E; ++I) { // We only add extract cost once for the same scalar. - if (!ExtractCostCalculated.insert(I->Scalar)) + if (!ExtractCostCalculated.insert(I->Scalar).second) + continue; + + // Uses by ephemeral values are free (because the ephemeral value will be + // removed prior to code generation, and so the extraction will be + // removed as well). + if (EphValues.count(I->User)) continue; VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth); @@ -1426,6 +1725,8 @@ int BoUpSLP::getTreeCost() { I->Lane); } + Cost += getSpillCost(); + DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n"); return Cost + ExtractCost; } @@ -1447,14 +1748,6 @@ int BoUpSLP::getGatherCost(ArrayRef VL) { return getGatherCost(VecTy); } -AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) { - if (StoreInst *SI = dyn_cast(I)) - return AA->getLocation(SI); - if (LoadInst *LI = dyn_cast(I)) - return AA->getLocation(LI); - return AliasAnalysis::Location(); -} - Value *BoUpSLP::getPointerOperand(Value *I) { if (LoadInst *LI = dyn_cast(I)) return LI->getPointerOperand(); @@ -1512,59 +1805,198 @@ bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) { return X == PtrSCEVB; } -Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) { - assert(Src->getParent() == Dst->getParent() && "Not the same BB"); - BasicBlock::iterator I = Src, E = Dst; - /// Scan all of the instruction from SRC to DST and check if - /// the source may alias. - for (++I; I != E; ++I) { - // Ignore store instructions that are marked as 'ignore'. - if (MemBarrierIgnoreList.count(I)) - continue; - if (Src->mayWriteToMemory()) /* Write */ { - if (!I->mayReadOrWriteMemory()) - continue; - } else /* Read */ { - if (!I->mayWriteToMemory()) - continue; - } - AliasAnalysis::Location A = getLocation(&*I); - AliasAnalysis::Location B = getLocation(Src); +// Reorder commutative operations in alternate shuffle if the resulting vectors +// are consecutive loads. This would allow us to vectorize the tree. +// If we have something like- +// load a[0] - load b[0] +// load b[1] + load a[1] +// load a[2] - load b[2] +// load a[3] + load b[3] +// Reordering the second load b[1] load a[1] would allow us to vectorize this +// code. +void BoUpSLP::reorderAltShuffleOperands(ArrayRef VL, + SmallVectorImpl &Left, + SmallVectorImpl &Right) { + + // Push left and right operands of binary operation into Left and Right + for (unsigned i = 0, e = VL.size(); i < e; ++i) { + Left.push_back(cast(VL[i])->getOperand(0)); + Right.push_back(cast(VL[i])->getOperand(1)); + } - if (!A.Ptr || !B.Ptr || AA->alias(A, B)) - return I; + // Reorder if we have a commutative operation and consecutive access + // are on either side of the alternate instructions. + for (unsigned j = 0; j < VL.size() - 1; ++j) { + if (LoadInst *L = dyn_cast(Left[j])) { + if (LoadInst *L1 = dyn_cast(Right[j + 1])) { + Instruction *VL1 = cast(VL[j]); + Instruction *VL2 = cast(VL[j + 1]); + if (isConsecutiveAccess(L, L1) && VL1->isCommutative()) { + std::swap(Left[j], Right[j]); + continue; + } else if (isConsecutiveAccess(L, L1) && VL2->isCommutative()) { + std::swap(Left[j + 1], Right[j + 1]); + continue; + } + // else unchanged + } + } + if (LoadInst *L = dyn_cast(Right[j])) { + if (LoadInst *L1 = dyn_cast(Left[j + 1])) { + Instruction *VL1 = cast(VL[j]); + Instruction *VL2 = cast(VL[j + 1]); + if (isConsecutiveAccess(L, L1) && VL1->isCommutative()) { + std::swap(Left[j], Right[j]); + continue; + } else if (isConsecutiveAccess(L, L1) && VL2->isCommutative()) { + std::swap(Left[j + 1], Right[j + 1]); + continue; + } + // else unchanged + } + } } - return nullptr; } -int BoUpSLP::getLastIndex(ArrayRef VL) { - BasicBlock *BB = cast(VL[0])->getParent(); - assert(BB == getSameBlock(VL) && "Invalid block"); - BlockNumbering &BN = getBlockNumbering(BB); +void BoUpSLP::reorderInputsAccordingToOpcode(ArrayRef VL, + SmallVectorImpl &Left, + SmallVectorImpl &Right) { - int MaxIdx = BN.getIndex(BB->getFirstNonPHI()); - for (unsigned i = 0, e = VL.size(); i < e; ++i) - MaxIdx = std::max(MaxIdx, BN.getIndex(cast(VL[i]))); - return MaxIdx; -} + SmallVector OrigLeft, OrigRight; -Instruction *BoUpSLP::getLastInstruction(ArrayRef VL) { - BasicBlock *BB = cast(VL[0])->getParent(); - assert(BB == getSameBlock(VL) && "Invalid block"); - BlockNumbering &BN = getBlockNumbering(BB); + bool AllSameOpcodeLeft = true; + bool AllSameOpcodeRight = true; + for (unsigned i = 0, e = VL.size(); i != e; ++i) { + Instruction *I = cast(VL[i]); + Value *VLeft = I->getOperand(0); + Value *VRight = I->getOperand(1); - int MaxIdx = BN.getIndex(cast(VL[0])); - for (unsigned i = 1, e = VL.size(); i < e; ++i) - MaxIdx = std::max(MaxIdx, BN.getIndex(cast(VL[i]))); - Instruction *I = BN.getInstruction(MaxIdx); - assert(I && "bad location"); - return I; + OrigLeft.push_back(VLeft); + OrigRight.push_back(VRight); + + Instruction *ILeft = dyn_cast(VLeft); + Instruction *IRight = dyn_cast(VRight); + + // Check whether all operands on one side have the same opcode. In this case + // we want to preserve the original order and not make things worse by + // reordering. + if (i && AllSameOpcodeLeft && ILeft) { + if (Instruction *PLeft = dyn_cast(OrigLeft[i - 1])) { + if (PLeft->getOpcode() != ILeft->getOpcode()) + AllSameOpcodeLeft = false; + } else + AllSameOpcodeLeft = false; + } + if (i && AllSameOpcodeRight && IRight) { + if (Instruction *PRight = dyn_cast(OrigRight[i - 1])) { + if (PRight->getOpcode() != IRight->getOpcode()) + AllSameOpcodeRight = false; + } else + AllSameOpcodeRight = false; + } + + // Sort two opcodes. In the code below we try to preserve the ability to use + // broadcast of values instead of individual inserts. + // vl1 = load + // vl2 = phi + // vr1 = load + // vr2 = vr2 + // = vl1 x vr1 + // = vl2 x vr2 + // If we just sorted according to opcode we would leave the first line in + // tact but we would swap vl2 with vr2 because opcode(phi) > opcode(load). + // = vl1 x vr1 + // = vr2 x vl2 + // Because vr2 and vr1 are from the same load we loose the opportunity of a + // broadcast for the packed right side in the backend: we have [vr1, vl2] + // instead of [vr1, vr2=vr1]. + if (ILeft && IRight) { + if (!i && ILeft->getOpcode() > IRight->getOpcode()) { + Left.push_back(IRight); + Right.push_back(ILeft); + } else if (i && ILeft->getOpcode() > IRight->getOpcode() && + Right[i - 1] != IRight) { + // Try not to destroy a broad cast for no apparent benefit. + Left.push_back(IRight); + Right.push_back(ILeft); + } else if (i && ILeft->getOpcode() == IRight->getOpcode() && + Right[i - 1] == ILeft) { + // Try preserve broadcasts. + Left.push_back(IRight); + Right.push_back(ILeft); + } else if (i && ILeft->getOpcode() == IRight->getOpcode() && + Left[i - 1] == IRight) { + // Try preserve broadcasts. + Left.push_back(IRight); + Right.push_back(ILeft); + } else { + Left.push_back(ILeft); + Right.push_back(IRight); + } + continue; + } + // One opcode, put the instruction on the right. + if (ILeft) { + Left.push_back(VRight); + Right.push_back(ILeft); + continue; + } + Left.push_back(VLeft); + Right.push_back(VRight); + } + + bool LeftBroadcast = isSplat(Left); + bool RightBroadcast = isSplat(Right); + + // If operands end up being broadcast return this operand order. + if (LeftBroadcast || RightBroadcast) + return; + + // Don't reorder if the operands where good to begin. + if (AllSameOpcodeRight || AllSameOpcodeLeft) { + Left = OrigLeft; + Right = OrigRight; + } + + // Finally check if we can get longer vectorizable chain by reordering + // without breaking the good operand order detected above. + // E.g. If we have something like- + // load a[0] load b[0] + // load b[1] load a[1] + // load a[2] load b[2] + // load a[3] load b[3] + // Reordering the second load b[1] load a[1] would allow us to vectorize + // this code and we still retain AllSameOpcode property. + // FIXME: This load reordering might break AllSameOpcode in some rare cases + // such as- + // add a[0],c[0] load b[0] + // add a[1],c[2] load b[1] + // b[2] load b[2] + // add a[3],c[3] load b[3] + for (unsigned j = 0; j < VL.size() - 1; ++j) { + if (LoadInst *L = dyn_cast(Left[j])) { + if (LoadInst *L1 = dyn_cast(Right[j + 1])) { + if (isConsecutiveAccess(L, L1)) { + std::swap(Left[j + 1], Right[j + 1]); + continue; + } + } + } + if (LoadInst *L = dyn_cast(Right[j])) { + if (LoadInst *L1 = dyn_cast(Left[j + 1])) { + if (isConsecutiveAccess(L, L1)) { + std::swap(Left[j + 1], Right[j + 1]); + continue; + } + } + } + // else unchanged + } } void BoUpSLP::setInsertPointAfterBundle(ArrayRef VL) { Instruction *VL0 = cast(VL[0]); - Instruction *LastInst = getLastInstruction(VL); - BasicBlock::iterator NextInst = LastInst; + BasicBlock::iterator NextInst = VL0; ++NextInst; Builder.SetInsertPoint(VL0->getParent(), NextInst); Builder.SetCurrentDebugLocation(VL0->getDebugLoc()); @@ -1647,6 +2079,7 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { setInsertPointAfterBundle(E->Scalars); return Gather(E->Scalars, VecTy); } + unsigned Opcode = getSameOpcode(E->Scalars); switch (Opcode) { @@ -1665,7 +2098,7 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { ValueList Operands; BasicBlock *IBB = PH->getIncomingBlock(i); - if (!VisitedBBs.insert(IBB)) { + if (!VisitedBBs.insert(IBB).second) { NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); continue; } @@ -1720,6 +2153,7 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { CastInst *CI = dyn_cast(VL0); Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); E->VectorizedValue = V; + ++NumVectorInstructions; return V; } case Instruction::FCmp: @@ -1746,6 +2180,7 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { V = Builder.CreateICmp(P0, L, R); E->VectorizedValue = V; + ++NumVectorInstructions; return V; } case Instruction::Select: { @@ -1767,6 +2202,7 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { Value *V = Builder.CreateSelect(Cond, True, False); E->VectorizedValue = V; + ++NumVectorInstructions; return V; } case Instruction::Add: @@ -1811,6 +2247,8 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { BinaryOperator *BinOp = cast(VL0); Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS); E->VectorizedValue = V; + propagateIRFlags(E->VectorizedValue, E->Scalars); + ++NumVectorInstructions; if (Instruction *I = dyn_cast(V)) return propagateMetadata(I, E->Scalars); @@ -1823,16 +2261,26 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { setInsertPointAfterBundle(E->Scalars); LoadInst *LI = cast(VL0); + Type *ScalarLoadTy = LI->getType(); unsigned AS = LI->getPointerAddressSpace(); Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(), VecTy->getPointerTo(AS)); + + // The pointer operand uses an in-tree scalar so we add the new BitCast to + // ExternalUses list to make sure that an extract will be generated in the + // future. + if (ScalarToTreeEntry.count(LI->getPointerOperand())) + ExternalUses.push_back( + ExternalUser(LI->getPointerOperand(), cast(VecPtr), 0)); + unsigned Alignment = LI->getAlignment(); LI = Builder.CreateLoad(VecPtr); if (!Alignment) - Alignment = DL->getABITypeAlignment(LI->getPointerOperand()->getType()); + Alignment = DL->getABITypeAlignment(ScalarLoadTy); LI->setAlignment(Alignment); E->VectorizedValue = LI; + ++NumVectorInstructions; return propagateMetadata(LI, E->Scalars); } case Instruction::Store: { @@ -1850,10 +2298,19 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(), VecTy->getPointerTo(AS)); StoreInst *S = Builder.CreateStore(VecValue, VecPtr); + + // The pointer operand uses an in-tree scalar so we add the new BitCast to + // ExternalUses list to make sure that an extract will be generated in the + // future. + if (ScalarToTreeEntry.count(SI->getPointerOperand())) + ExternalUses.push_back( + ExternalUser(SI->getPointerOperand(), cast(VecPtr), 0)); + if (!Alignment) - Alignment = DL->getABITypeAlignment(SI->getPointerOperand()->getType()); + Alignment = DL->getABITypeAlignment(SI->getValueOperand()->getType()); S->setAlignment(Alignment); E->VectorizedValue = S; + ++NumVectorInstructions; return propagateMetadata(S, E->Scalars); } case Instruction::GetElementPtr: { @@ -1878,6 +2335,7 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { Value *V = Builder.CreateGEP(Op0, OpVecs); E->VectorizedValue = V; + ++NumVectorInstructions; if (Instruction *I = dyn_cast(V)) return propagateMetadata(I, E->Scalars); @@ -1889,6 +2347,7 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { setInsertPointAfterBundle(E->Scalars); Function *FI; Intrinsic::ID IID = Intrinsic::not_intrinsic; + Value *ScalarArg = nullptr; if (CI && (FI = CI->getCalledFunction())) { IID = (Intrinsic::ID) FI->getIntrinsicID(); } @@ -1899,6 +2358,7 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { // a scalar. This argument should not be vectorized. if (hasVectorInstrinsicScalarOpd(IID, 1) && j == 1) { CallInst *CEI = cast(E->Scalars[0]); + ScalarArg = CEI->getArgOperand(j); OpVecs.push_back(CEI->getArgOperand(j)); continue; } @@ -1917,15 +2377,21 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) }; Function *CF = Intrinsic::getDeclaration(M, ID, Tys); Value *V = Builder.CreateCall(CF, OpVecs); + + // The scalar argument uses an in-tree scalar so we add the new vectorized + // call to ExternalUses list to make sure that an extract will be + // generated in the future. + if (ScalarArg && ScalarToTreeEntry.count(ScalarArg)) + ExternalUses.push_back(ExternalUser(ScalarArg, cast(V), 0)); + E->VectorizedValue = V; + ++NumVectorInstructions; return V; } case Instruction::ShuffleVector: { ValueList LHSVL, RHSVL; - for (int i = 0, e = E->Scalars.size(); i < e; ++i) { - LHSVL.push_back(cast(E->Scalars[i])->getOperand(0)); - RHSVL.push_back(cast(E->Scalars[i])->getOperand(1)); - } + assert(isa(VL0) && "Invalid Shuffle Vector Operand"); + reorderAltShuffleOperands(E->Scalars, LHSVL, RHSVL); setInsertPointAfterBundle(E->Scalars); Value *LHS = vectorizeTree(LHSVL); @@ -1943,21 +2409,29 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { BinaryOperator *BinOp1 = cast(VL1); Value *V1 = Builder.CreateBinOp(BinOp1->getOpcode(), LHS, RHS); - // Create appropriate shuffle to take alternative operations from - // the vector. - std::vector Mask(E->Scalars.size()); + // Create shuffle to take alternate operations from the vector. + // Also, gather up odd and even scalar ops to propagate IR flags to + // each vector operation. + ValueList OddScalars, EvenScalars; unsigned e = E->Scalars.size(); + SmallVector Mask(e); for (unsigned i = 0; i < e; ++i) { - if (i & 1) + if (i & 1) { Mask[i] = Builder.getInt32(e + i); - else + OddScalars.push_back(E->Scalars[i]); + } else { Mask[i] = Builder.getInt32(i); + EvenScalars.push_back(E->Scalars[i]); + } } Value *ShuffleMask = ConstantVector::get(Mask); + propagateIRFlags(V0, EvenScalars); + propagateIRFlags(V1, OddScalars); Value *V = Builder.CreateShuffleVector(V0, V1, ShuffleMask); E->VectorizedValue = V; + ++NumVectorInstructions; if (Instruction *I = dyn_cast(V)) return propagateMetadata(I, E->Scalars); @@ -1970,6 +2444,12 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { } Value *BoUpSLP::vectorizeTree() { + + // All blocks must be scheduled before any instructions are inserted. + for (auto &BSIter : BlocksSchedules) { + scheduleBlock(BSIter.second.get()); + } + Builder.SetInsertPoint(F->getEntryBlock().begin()); vectorizeTree(&VectorizableTree[0]); @@ -2054,13 +2534,10 @@ Value *BoUpSLP::vectorizeTree() { Scalar->replaceAllUsesWith(Undef); } DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n"); - cast(Scalar)->eraseFromParent(); + eraseInstruction(cast(Scalar)); } } - for (auto &BN : BlocksNumbers) - BN.second.forget(); - Builder.ClearInsertionPoint(); return VectorizableTree[0].VectorizedValue; @@ -2139,7 +2616,7 @@ void BoUpSLP::optimizeGatherSequence() { if (In->isIdenticalTo(*v) && DT->dominates((*v)->getParent(), In->getParent())) { In->replaceAllUsesWith(*v); - In->eraseFromParent(); + eraseInstruction(In); In = nullptr; break; } @@ -2154,6 +2631,388 @@ void BoUpSLP::optimizeGatherSequence() { GatherSeq.clear(); } +// Groups the instructions to a bundle (which is then a single scheduling entity) +// and schedules instructions until the bundle gets ready. +bool BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef VL, + BoUpSLP *SLP) { + if (isa(VL[0])) + return true; + + // Initialize the instruction bundle. + Instruction *OldScheduleEnd = ScheduleEnd; + ScheduleData *PrevInBundle = nullptr; + ScheduleData *Bundle = nullptr; + bool ReSchedule = false; + DEBUG(dbgs() << "SLP: bundle: " << *VL[0] << "\n"); + for (Value *V : VL) { + extendSchedulingRegion(V); + ScheduleData *BundleMember = getScheduleData(V); + assert(BundleMember && + "no ScheduleData for bundle member (maybe not in same basic block)"); + if (BundleMember->IsScheduled) { + // A bundle member was scheduled as single instruction before and now + // needs to be scheduled as part of the bundle. We just get rid of the + // existing schedule. + DEBUG(dbgs() << "SLP: reset schedule because " << *BundleMember + << " was already scheduled\n"); + ReSchedule = true; + } + assert(BundleMember->isSchedulingEntity() && + "bundle member already part of other bundle"); + if (PrevInBundle) { + PrevInBundle->NextInBundle = BundleMember; + } else { + Bundle = BundleMember; + } + BundleMember->UnscheduledDepsInBundle = 0; + Bundle->UnscheduledDepsInBundle += BundleMember->UnscheduledDeps; + + // Group the instructions to a bundle. + BundleMember->FirstInBundle = Bundle; + PrevInBundle = BundleMember; + } + if (ScheduleEnd != OldScheduleEnd) { + // The scheduling region got new instructions at the lower end (or it is a + // new region for the first bundle). This makes it necessary to + // recalculate all dependencies. + // It is seldom that this needs to be done a second time after adding the + // initial bundle to the region. + for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { + ScheduleData *SD = getScheduleData(I); + SD->clearDependencies(); + } + ReSchedule = true; + } + if (ReSchedule) { + resetSchedule(); + initialFillReadyList(ReadyInsts); + } + + DEBUG(dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " + << BB->getName() << "\n"); + + calculateDependencies(Bundle, true, SLP); + + // Now try to schedule the new bundle. As soon as the bundle is "ready" it + // means that there are no cyclic dependencies and we can schedule it. + // Note that's important that we don't "schedule" the bundle yet (see + // cancelScheduling). + while (!Bundle->isReady() && !ReadyInsts.empty()) { + + ScheduleData *pickedSD = ReadyInsts.back(); + ReadyInsts.pop_back(); + + if (pickedSD->isSchedulingEntity() && pickedSD->isReady()) { + schedule(pickedSD, ReadyInsts); + } + } + return Bundle->isReady(); +} + +void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef VL) { + if (isa(VL[0])) + return; + + ScheduleData *Bundle = getScheduleData(VL[0]); + DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n"); + assert(!Bundle->IsScheduled && + "Can't cancel bundle which is already scheduled"); + assert(Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && + "tried to unbundle something which is not a bundle"); + + // Un-bundle: make single instructions out of the bundle. + ScheduleData *BundleMember = Bundle; + while (BundleMember) { + assert(BundleMember->FirstInBundle == Bundle && "corrupt bundle links"); + BundleMember->FirstInBundle = BundleMember; + ScheduleData *Next = BundleMember->NextInBundle; + BundleMember->NextInBundle = nullptr; + BundleMember->UnscheduledDepsInBundle = BundleMember->UnscheduledDeps; + if (BundleMember->UnscheduledDepsInBundle == 0) { + ReadyInsts.insert(BundleMember); + } + BundleMember = Next; + } +} + +void BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V) { + if (getScheduleData(V)) + return; + Instruction *I = dyn_cast(V); + assert(I && "bundle member must be an instruction"); + assert(!isa(I) && "phi nodes don't need to be scheduled"); + if (!ScheduleStart) { + // It's the first instruction in the new region. + initScheduleData(I, I->getNextNode(), nullptr, nullptr); + ScheduleStart = I; + ScheduleEnd = I->getNextNode(); + assert(ScheduleEnd && "tried to vectorize a TerminatorInst?"); + DEBUG(dbgs() << "SLP: initialize schedule region to " << *I << "\n"); + return; + } + // Search up and down at the same time, because we don't know if the new + // instruction is above or below the existing scheduling region. + BasicBlock::reverse_iterator UpIter(ScheduleStart); + BasicBlock::reverse_iterator UpperEnd = BB->rend(); + BasicBlock::iterator DownIter(ScheduleEnd); + BasicBlock::iterator LowerEnd = BB->end(); + for (;;) { + if (UpIter != UpperEnd) { + if (&*UpIter == I) { + initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion); + ScheduleStart = I; + DEBUG(dbgs() << "SLP: extend schedule region start to " << *I << "\n"); + return; + } + UpIter++; + } + if (DownIter != LowerEnd) { + if (&*DownIter == I) { + initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion, + nullptr); + ScheduleEnd = I->getNextNode(); + assert(ScheduleEnd && "tried to vectorize a TerminatorInst?"); + DEBUG(dbgs() << "SLP: extend schedule region end to " << *I << "\n"); + return; + } + DownIter++; + } + assert((UpIter != UpperEnd || DownIter != LowerEnd) && + "instruction not found in block"); + } +} + +void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI, + Instruction *ToI, + ScheduleData *PrevLoadStore, + ScheduleData *NextLoadStore) { + ScheduleData *CurrentLoadStore = PrevLoadStore; + for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) { + ScheduleData *SD = ScheduleDataMap[I]; + if (!SD) { + // Allocate a new ScheduleData for the instruction. + if (ChunkPos >= ChunkSize) { + ScheduleDataChunks.push_back( + llvm::make_unique(ChunkSize)); + ChunkPos = 0; + } + SD = &(ScheduleDataChunks.back()[ChunkPos++]); + ScheduleDataMap[I] = SD; + SD->Inst = I; + } + assert(!isInSchedulingRegion(SD) && + "new ScheduleData already in scheduling region"); + SD->init(SchedulingRegionID); + + if (I->mayReadOrWriteMemory()) { + // Update the linked list of memory accessing instructions. + if (CurrentLoadStore) { + CurrentLoadStore->NextLoadStore = SD; + } else { + FirstLoadStoreInRegion = SD; + } + CurrentLoadStore = SD; + } + } + if (NextLoadStore) { + if (CurrentLoadStore) + CurrentLoadStore->NextLoadStore = NextLoadStore; + } else { + LastLoadStoreInRegion = CurrentLoadStore; + } +} + +void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD, + bool InsertInReadyList, + BoUpSLP *SLP) { + assert(SD->isSchedulingEntity()); + + SmallVector WorkList; + WorkList.push_back(SD); + + while (!WorkList.empty()) { + ScheduleData *SD = WorkList.back(); + WorkList.pop_back(); + + ScheduleData *BundleMember = SD; + while (BundleMember) { + assert(isInSchedulingRegion(BundleMember)); + if (!BundleMember->hasValidDependencies()) { + + DEBUG(dbgs() << "SLP: update deps of " << *BundleMember << "\n"); + BundleMember->Dependencies = 0; + BundleMember->resetUnscheduledDeps(); + + // Handle def-use chain dependencies. + for (User *U : BundleMember->Inst->users()) { + if (isa(U)) { + ScheduleData *UseSD = getScheduleData(U); + if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) { + BundleMember->Dependencies++; + ScheduleData *DestBundle = UseSD->FirstInBundle; + if (!DestBundle->IsScheduled) { + BundleMember->incrementUnscheduledDeps(1); + } + if (!DestBundle->hasValidDependencies()) { + WorkList.push_back(DestBundle); + } + } + } else { + // I'm not sure if this can ever happen. But we need to be safe. + // This lets the instruction/bundle never be scheduled and eventally + // disable vectorization. + BundleMember->Dependencies++; + BundleMember->incrementUnscheduledDeps(1); + } + } + + // Handle the memory dependencies. + ScheduleData *DepDest = BundleMember->NextLoadStore; + if (DepDest) { + Instruction *SrcInst = BundleMember->Inst; + AliasAnalysis::Location SrcLoc = getLocation(SrcInst, SLP->AA); + bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory(); + unsigned numAliased = 0; + unsigned DistToSrc = 1; + + while (DepDest) { + assert(isInSchedulingRegion(DepDest)); + + // We have two limits to reduce the complexity: + // 1) AliasedCheckLimit: It's a small limit to reduce calls to + // SLP->isAliased (which is the expensive part in this loop). + // 2) MaxMemDepDistance: It's for very large blocks and it aborts + // the whole loop (even if the loop is fast, it's quadratic). + // It's important for the loop break condition (see below) to + // check this limit even between two read-only instructions. + if (DistToSrc >= MaxMemDepDistance || + ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) && + (numAliased >= AliasedCheckLimit || + SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) { + + // We increment the counter only if the locations are aliased + // (instead of counting all alias checks). This gives a better + // balance between reduced runtime and accurate dependencies. + numAliased++; + + DepDest->MemoryDependencies.push_back(BundleMember); + BundleMember->Dependencies++; + ScheduleData *DestBundle = DepDest->FirstInBundle; + if (!DestBundle->IsScheduled) { + BundleMember->incrementUnscheduledDeps(1); + } + if (!DestBundle->hasValidDependencies()) { + WorkList.push_back(DestBundle); + } + } + DepDest = DepDest->NextLoadStore; + + // Example, explaining the loop break condition: Let's assume our + // starting instruction is i0 and MaxMemDepDistance = 3. + // + // +--------v--v--v + // i0,i1,i2,i3,i4,i5,i6,i7,i8 + // +--------^--^--^ + // + // MaxMemDepDistance let us stop alias-checking at i3 and we add + // dependencies from i0 to i3,i4,.. (even if they are not aliased). + // Previously we already added dependencies from i3 to i6,i7,i8 + // (because of MaxMemDepDistance). As we added a dependency from + // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8 + // and we can abort this loop at i6. + if (DistToSrc >= 2 * MaxMemDepDistance) + break; + DistToSrc++; + } + } + } + BundleMember = BundleMember->NextInBundle; + } + if (InsertInReadyList && SD->isReady()) { + ReadyInsts.push_back(SD); + DEBUG(dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n"); + } + } +} + +void BoUpSLP::BlockScheduling::resetSchedule() { + assert(ScheduleStart && + "tried to reset schedule on block which has not been scheduled"); + for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { + ScheduleData *SD = getScheduleData(I); + assert(isInSchedulingRegion(SD)); + SD->IsScheduled = false; + SD->resetUnscheduledDeps(); + } + ReadyInsts.clear(); +} + +void BoUpSLP::scheduleBlock(BlockScheduling *BS) { + + if (!BS->ScheduleStart) + return; + + DEBUG(dbgs() << "SLP: schedule block " << BS->BB->getName() << "\n"); + + BS->resetSchedule(); + + // For the real scheduling we use a more sophisticated ready-list: it is + // sorted by the original instruction location. This lets the final schedule + // be as close as possible to the original instruction order. + struct ScheduleDataCompare { + bool operator()(ScheduleData *SD1, ScheduleData *SD2) { + return SD2->SchedulingPriority < SD1->SchedulingPriority; + } + }; + std::set ReadyInsts; + + // Ensure that all depencency data is updated and fill the ready-list with + // initial instructions. + int Idx = 0; + int NumToSchedule = 0; + for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; + I = I->getNextNode()) { + ScheduleData *SD = BS->getScheduleData(I); + assert( + SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) && + "scheduler and vectorizer have different opinion on what is a bundle"); + SD->FirstInBundle->SchedulingPriority = Idx++; + if (SD->isSchedulingEntity()) { + BS->calculateDependencies(SD, false, this); + NumToSchedule++; + } + } + BS->initialFillReadyList(ReadyInsts); + + Instruction *LastScheduledInst = BS->ScheduleEnd; + + // Do the "real" scheduling. + while (!ReadyInsts.empty()) { + ScheduleData *picked = *ReadyInsts.begin(); + ReadyInsts.erase(ReadyInsts.begin()); + + // Move the scheduled instruction(s) to their dedicated places, if not + // there yet. + ScheduleData *BundleMember = picked; + while (BundleMember) { + Instruction *pickedInst = BundleMember->Inst; + if (LastScheduledInst->getNextNode() != pickedInst) { + BS->BB->getInstList().remove(pickedInst); + BS->BB->getInstList().insert(LastScheduledInst, pickedInst); + } + LastScheduledInst = pickedInst; + BundleMember = BundleMember->NextInBundle; + } + + BS->schedule(picked, ReadyInsts); + NumToSchedule--; + } + assert(NumToSchedule == 0 && "could not schedule all instructions"); + + // Avoid duplicate scheduling of the block. + BS->ScheduleStart = nullptr; +} + /// The SLPVectorizer Pass. struct SLPVectorizer : public FunctionPass { typedef SmallVector StoreList; @@ -2173,6 +3032,7 @@ struct SLPVectorizer : public FunctionPass { AliasAnalysis *AA; LoopInfo *LI; DominatorTree *DT; + AssumptionCache *AC; bool runOnFunction(Function &F) override { if (skipOptnoneFunction(F)) @@ -2182,10 +3042,12 @@ struct SLPVectorizer : public FunctionPass { DataLayoutPass *DLP = getAnalysisIfAvailable(); DL = DLP ? &DLP->getDataLayout() : nullptr; TTI = &getAnalysis(); - TLI = getAnalysisIfAvailable(); + auto *TLIP = getAnalysisIfAvailable(); + TLI = TLIP ? &TLIP->getTLI() : nullptr; AA = &getAnalysis(); - LI = &getAnalysis(); + LI = &getAnalysis().getLoopInfo(); DT = &getAnalysis().getDomTree(); + AC = &getAnalysis().getAssumptionCache(F); StoreRefs.clear(); bool Changed = false; @@ -2208,7 +3070,10 @@ struct SLPVectorizer : public FunctionPass { // Use the bottom up slp vectorizer to construct chains that start with // store instructions. - BoUpSLP R(&F, SE, DL, TTI, TLI, AA, LI, DT); + BoUpSLP R(&F, SE, DL, TTI, TLI, AA, LI, DT, AC); + + // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to + // delete instructions. // Scan the blocks in the function in post order. for (po_iterator it = po_begin(&F.getEntryBlock()), @@ -2235,12 +3100,13 @@ struct SLPVectorizer : public FunctionPass { void getAnalysisUsage(AnalysisUsage &AU) const override { FunctionPass::getAnalysisUsage(AU); + AU.addRequired(); AU.addRequired(); AU.addRequired(); AU.addRequired(); - AU.addRequired(); + AU.addRequired(); AU.addRequired(); - AU.addPreserved(); + AU.addPreserved(); AU.addPreserved(); AU.setPreservesCFG(); } @@ -2551,11 +3417,9 @@ bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) { BinaryOperator *B0 = dyn_cast(B->getOperand(0)); BinaryOperator *B1 = dyn_cast(B->getOperand(1)); if (tryToVectorizePair(A, B0, R)) { - B->moveBefore(V); return true; } if (tryToVectorizePair(A, B1, R)) { - B->moveBefore(V); return true; } } @@ -2565,11 +3429,9 @@ bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) { BinaryOperator *A0 = dyn_cast(A->getOperand(0)); BinaryOperator *A1 = dyn_cast(A->getOperand(1)); if (tryToVectorizePair(A0, B, R)) { - A->moveBefore(V); return true; } if (tryToVectorizePair(A1, B, R)) { - A->moveBefore(V); return true; } } @@ -2765,8 +3627,7 @@ public: unsigned i = 0; for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) { - ArrayRef ValsToReduce(&ReducedVals[i], ReduxWidth); - V.buildTree(ValsToReduce, ReductionOps); + V.buildTree(makeArrayRef(&ReducedVals[i], ReduxWidth), ReductionOps); // Estimate cost. int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]); @@ -2844,11 +3705,10 @@ private: /// \brief Emit a horizontal reduction of the vectorized value. Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) { assert(VectorizedValue && "Need to have a vectorized tree node"); - Instruction *ValToReduce = dyn_cast(VectorizedValue); assert(isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now"); - Value *TmpVec = ValToReduce; + Value *TmpVec = VectorizedValue; for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) { if (IsPairwiseReduction) { Value *LeftMask = @@ -2958,8 +3818,7 @@ bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { // Try to vectorize them. unsigned NumElts = (SameTypeIt - IncIt); DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n"); - if (NumElts > 1 && - tryToVectorizeList(ArrayRef(IncIt, NumElts), R)) { + if (NumElts > 1 && tryToVectorizeList(makeArrayRef(IncIt, NumElts), R)) { // Success start over because instructions might have been changed. HaveVectorizedPhiNodes = true; Changed = true; @@ -2975,7 +3834,7 @@ bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) { // We may go through BB multiple times so skip the one we have checked. - if (!VisitedInstrs.insert(it)) + if (!VisitedInstrs.insert(it).second) continue; if (isa(it)) @@ -3039,6 +3898,21 @@ bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { } } + // Try to vectorize horizontal reductions feeding into a return. + if (ReturnInst *RI = dyn_cast(it)) + if (RI->getNumOperands() != 0) + if (BinaryOperator *BinOp = + dyn_cast(RI->getOperand(0))) { + DEBUG(dbgs() << "SLP: Found a return to vectorize.\n"); + if (tryToVectorizePair(BinOp->getOperand(0), + BinOp->getOperand(1), R)) { + Changed = true; + it = BB->begin(); + e = BB->end(); + continue; + } + } + // Try to vectorize trees that start at compare instructions. if (CmpInst *CI = dyn_cast(it)) { if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) { @@ -3101,8 +3975,8 @@ bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) { // Process the stores in chunks of 16. for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) { unsigned Len = std::min(CE - CI, 16); - ArrayRef Chunk(&it->second[CI], Len); - Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R); + Changed |= vectorizeStores(makeArrayRef(&it->second[CI], Len), + -SLPCostThreshold, R); } } return Changed; @@ -3115,6 +3989,7 @@ static const char lv_name[] = "SLP Vectorizer"; INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false) INITIALIZE_AG_DEPENDENCY(AliasAnalysis) INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) +INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) INITIALIZE_PASS_DEPENDENCY(LoopSimplify) INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)